Anthropic Unveils FlashProve to Cryptographically Verify AI Model Outputs

Image: TechCrunch · Source

Anthropic introduces FlashProve, an open-source cryptographic system that enables verification of whether AI responses genuinely originate from specific models, enhancing trust and security in AI usage.

Anthropic has developed FlashProve, an open-source system designed to cryptographically verify the identity of AI models producing specific outputs. Published in November 2026, this research tackles a notable challenge in AI interactions: users currently cannot be certain that a response comes from the exact AI model claimed by the provider. FlashProve aims to solve this by enabling "provable inference," providing cryptographic proofs that a particular response was generated by a designated AI model.

The problem arises because organizations or malicious actors might substitute one model for another without users' knowledge—potentially switching to cheaper models or inflating performance claims. FlashProve draws parallels to cryptographic email signatures, as it attaches proofs to AI outputs that can be independently and quickly verified, even on devices with modest computational power.

Technically, FlashProve uses zero-knowledge succinct arguments of knowledge (zkSNARKs), a type of cryptographic protocol that allows proving computational statements without revealing sensitive data such as the model's internal weights. Importantly, the system requires AI labs to publish a "fingerprint" or commitment to their model weights. These fingerprints do not disclose the model weights themselves, thus preserving intellectual property while enabling verification.

Performance-wise, for a large Mixture-of-Experts language model like Llama 4 Maverick (400 billion parameters, 4-bit quantization), a proof can be generated at about 110 tokens per second on a single Nvidia H100 GPU, with a proof size of roughly 19 MB per 1,024 tokens. Verification on a typical multi-core CPU takes only a few seconds, requiring no specialized hardware. While producing proofs is computationally intensive—about 70 to 95 times longer than running inference alone—the verification process is lightweight.

Potential applications include validating that outputs or analyses genuinely originate from advanced or approved models, crucial for clients auditing AI services or regulators enforcing compliance. It also guards against tampering with inference servers, as attackers cannot forge valid proofs without the correct model weights.

Anthropic released FlashProve's code, formal verification proofs written in the Lean proof assistant, and extensive technical documentation under an open-source Apache 2.0 license, inviting the community to build upon their work. They emphasize that FlashProve is a research prototype, not yet deployed in production, with ongoing investigations into further improvements.

By enabling cryptographic attestation of AI outputs without compromising proprietary model details, FlashProve represents a significant step toward building trustworthy AI systems that can be reliably audited and verified by users, evaluators, and regulators alike.